Return the renderer for an arena, or ``None`` if none exists yet. Imports are lazy so pulling in one arena's replay code never drags in the other

Load metadata.json from log directory and return Metadata object

Parse a specific trajectory file Args: player_name: Name of the player round_num: Round number load_diffs

Calculate Elo for max rounds from 1 to max_rounds and plot the results.

Save figure to PDF.

Set cache value

Create engine log file with final scores. Args: scores: List of (player_id, final_money) tuples

Evaluate a matrix between two players using manual thread management.

Get the diff between the round and initial version (round 0). If incremental is True, get the diff between the round and the previous round.

Save metadata to matrix.json.

Save plot in both PDF and PNG formats.

Save plot in both PDF and PNG formats.

Local scratch dir for staging `docker cp` transfers. Defaults to the system temp dir. Override with CODECLASH_TMPDIR (e.g. on AWS Batch, where the

JSON for inlining inside a <script> tag (guard against ``</script>``).

Git tag the codebase at the given round.

Add a file handler to the root logger to capture all log messages. Returns: The FileHandler that was added (for later cleanup).

Aggregate all llm_as_judge.json results from the input directory into a DataFrame. Args: input_dir: Directory containing llm_as_j

Calculate filename redundancy over rounds for each player-tournament.

One-step-ahead residuals for AR(p) on logit(scores) with stationary means. scores: (T, R) in (0,1) gammas: (p,) mus: (T,) Return

Convert Bradley-Terry strength to Elo rating. Formula: R_i = R_0 + (β/ln(10)) * s_i where β = 400 (ELO_SLOPE), R_0 = 1200 (ELO_BASE)

Walk logs and build model -> list[list[int]] for rounds 1..ROUNDS. Returns a dict where each model maps to a list of ROUNDS lists, where each

Walk logs and build model -> list[list[int]] for rounds 1..ROUNDS. Returns a dict where each model maps to a list of ROUNDS lists, where each

Calculate how often the winner changes between consecutive rounds.

Calculate the p-value for the statistical significance of the winner. Input: scores as a dictionary of player names and number of games won.

Calculate the number of throwaway files at root and elsewhere. A throwaway file is one that is created in round i < threshold_round and never use

Calculate winner's share of total points for each round in a tournament.

Close and remove file handlers to prevent file descriptor leaks.

Clear all cache entries

Copy files from one Docker container to another via a temporary local directory. Be extremely careful with trailing slashes in src_path and

* Create the line count chart for a specific file

Create all three plots.

Find every playable game under a tournament folder. Handles three layouts: ``rounds/round_<R>.tar.gz`` archives (the common case), uncompress

Extract file path from a git diff block.

Create an agent with environment and game context.

Create a dummy agent that does nothing.

Create a game context for an agent.

Get cached value or compute it, preventing concurrent computation Args: key: Cache key compute_fn: Function to call i

Load all cached results from the data file.

Parse overall game metadata

Plot average lines changed per round for each model and save PNG. file_types: iterable of file-type keys to include (e.g. ["README_agent.md"] or

Plot heatmap showing how model consistency varies by opponent. Answers questions 2a (round 1) and 2b (round 15).

Plot win counts stratified by category (game or model).

Plot normalized scores stratified by category (game or model).

Should be called after we called the run method.

Should be called before we call the run method.

* Prepare chart data for a specific file

Common logic for rendering game viewer pages with timeout protection

Main execution function that runs all rounds.

Sync local logs to S3 bucket. Args: local_output_dir: Local directory containing logs to sync logger: Logger instance for debug o

Set the logs directory directly

* Setup the file dropdown change handler

Split a git diff into separate diffs for each file. Args: diff: Git diff text containing potentially multiple files Returns: